Research map / Image generation / Text-to-image diffusion
StableRep: Synthetic Images from Text-to-Image Models Make Strong Visual Representation Learners
Yonglong Tian, Lijie Fan, Phillip Isola, Huiwen Chang, Dilip Krishnan
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train StableRep with synthetic images and multi-positive same-caption contrast | StableRep vs CLIP trained on real images · ImageNet linear probe, CC12M | accuracy: 70.3 -> 73.5 | Improved |
| Train StableRep with synthetic images and multi-positive same-caption contrast | StableRep vs CLIP trained on real images · ImageNet linear probe, RedCaps | accuracy: 71.9 -> 74.5 | Improved |
| Replace real images with synthetic images for SimCLR pretraining | SimCLR · ImageNet linear probe, CC12M | accuracy: 61.5 -> 63.7 | Improved |
| Replace real images with synthetic images for CLIP pretraining | CLIP · ImageNet linear probe, CC12M | accuracy: 70.3 -> 67.8 | Hurt |
| Generate multiple images per caption instead of one (ablation) | StableRep · ImageNet linear probe, CC3M, fixed image-generation budget | accuracy: 61.2 -> 66.2 | Improved |
| Increase images per caption in a batch from 1 to 8 (ablation) | StableRep · ImageNet linear probe, CC3M, fixed batch size | accuracy: 60.5 -> 69.8 | Improved |
| Use smaller generation guidance scales w in {2,3} instead of w in {8,10} (ablation) | StableRep · ImageNet linear probe, CC12M | accuracy: 70.8 -> 72.8 | Improved |
| Scale the backbone from ViT-B/16 to ViT-L/16 (ablation) | StableRep · ImageNet linear probe, CC12M | accuracy: 72.8 -> 74.7 | Improved |
| Train StableRep longer, from 35 to 105 epochs (ablation) | StableRep · ImageNet linear probe, CC12M | accuracy: 72.8 -> 75.7 | Improved |
| Add random downsample augmentation (ablation) | StableRep · Fine-grained linear probing average, CC12M | accuracy: 82.2 -> 83.7 | Improved |
| Add a non-affine BatchNorm layer to the classifier (ablation) | StableRep trained for 35 epochs · ImageNet linear probe | accuracy: 71.4 -> 73.5 | Improved |
| Add a non-affine BatchNorm layer to the classifier (ablation) | StableRep trained for 105 epochs · ImageNet linear probe | accuracy: 75.4 -> 75.2 | No clear effect |
Closest papers by meaning
- Learning Vision from Models Rivals Learning Vision from Data (CVPR 2024 · similarity 0.91)
- The Unmet Promise of Synthetic Training Images: Using Retrieved Real Images Performs Better (NeurIPS 2024 · similarity 0.90)
- Image Captioning with Multi-Context Synthetic Data (AAAI 2024 · similarity 0.88)
- FakeInversion: Learning to Detect Images from Unseen Text-to-Image Models by Inverting Stable Diffusion (CVPR 2024 · similarity 0.88)
- Enhancing Vision-Language Compositional Understanding with Multimodal Synthetic Data (CVPR 2025 · similarity 0.88)
- Scaling Laws of Synthetic Images for Model Training ... for Now (CVPR 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
